Safe AI Agents

Research

The Safe AI Agents (SAIA) Initiative conducts interdisciplinary research focused on the safety, governance, alignment, and operational trustworthiness of AI agents. As AI systems become increasingly autonomous and integrated into enterprise and societal environments, SAIA explores both the theoretical foundations and practical frameworks required to enable the responsible development, deployment, and operation of safe AI agents at scale.

The initiative combines perspectives from AI research, engineering, governance, public deliberation, and operational systems to address emerging challenges associated with autonomous and multi-agent systems.

Research Interests

Current research interests of the SAIA Initiative include:

  • AI agent governance and operational safety
  • Human-in-the-loop (HITL) systems and human oversight
  • Reinforcement learning from human feedback (RLHF) for AI agents
  • Progressive autonomy and staged automation
  • Transparency, accountability, and trustworthiness in AI systems
  • Policy-driven control and governance frameworks
  • Agent observability, monitoring, and auditability
  • Multi-agent coordination and controlled autonomy
  • AI alignment and human value alignment
  • Runtime safety mechanisms and fallback systems
  • Enterprise deployment and scaling of AI agents
  • Public deliberation and societal impacts of autonomous AI systems
  • AI Governance frameworks and regulatory alignment (ISO/IEC 42001, EU AI Act, NIST AI RMF)
  • Operational architectures for scalable and safe AI agents orchestration
  • Enhancing contextual understanding and coordination in multi-agent workflows
  • Investigating how computer vision inputs and validated human decisions can inform the progressive granting of autonomous decision authority to AI agents
  • Agent governance in drones, robotics, embodied AI, and industrial inspection applications

Current Research

The SAIA Initiative actively conducts research on frameworks, architectures, and operational models for safe AI agents. Current work includes research on human oversight systems, progressive autonomy, governance frameworks, reinforcement learning from human feedback (RLHF), policy-driven agent control, and operational platforms for trustworthy AI agents.

The initiative also collaborates with industry partners and researchers to explore practical approaches for deploying AI agents safely and responsibly in real-world environments.

  • Toward Safe and Responsible AI Agents: A Three-Pillar Model for Transparency, Accountability, and Trustworthiness (https://arxiv.org/abs/2601.06223)
  • A Decoupled Human-in-the-Loop System for Controlled Autonomy in Agentic Workflows (https://arxiv.org/abs/2604.23049)
  • RLHF for Progressive, Governed Autonomy in Agentic Systems (submitted for publication)
  • Policy-Governed Progressive Autonomy for Computer Vision Agents in Vision Critical Applications (submitted for publication)
  • A Three-Pillar Governance Framework for Safe and Trustworthy AI Agents (accepted for publication - IEEE AGCS 2026)